Karpathy’s LLM Wiki – How Teachers Can Use It

Karpathy's LLM Wiki - How Teachers Can Use It

Knowledge Systems 💡 Concept Tutorial ↺ 11 min Apr 7, 2026

What You’ll Learn

Andrej Karpathy’s LLM Wiki idea went viral for a reason — it solves a real problem: years of digital files scattered across your hard drive with no easy way to search or reference them. In this tutorial, you’ll understand exactly what the LLM Wiki is, why it’s a powerful personal tool, and — critically — what it can and can’t do for your students. You’ll also see the two-layer model James uses at trainingsites.io that solves this for educators.

What Is Karpathy’s LLM Wiki?

The idea is simple: take all the “digital exhaust” sitting on your computer — videos, PDFs, Google Docs, transcripts, notes, images — and use an AI like Claude to convert it into a structured, visual knowledge base called a wiki.

https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f

The tool that makes it visual is Obsidian — free software that displays your notes and documents as an interconnected graph. Think of it like a map of everything you know, where related topics are connected by lines.

“If we could have Obsidian running on our local computer, it would be a really good way for us to see it — but also to have a wiki created locally with large language models like Claude to create a personal wiki.” — James

The three components of the setup are: a raw folder (where you drop all your existing files), a Claude-powered ingest process (which reads each file and converts it into a wiki page with title, topics, sources, and summaries), and the Obsidian vault (where you browse and query everything visually).

What It Looks Like in Practice

James built a vault from 607 YouTube video folders. Each folder contained a mix of raw video files, PDFs, Google Docs, transcripts, and images. After running the ingest, Claude organized them into structured wiki pages with:

  • Title and content type
  • Topics and categories
  • Source files referenced
  • Summaries of what was covered
  • Cross-references to related videos

The graph view in Obsidian shows every wiki page as a dot, with connecting lines to related content. The biggest dots — most referenced — were topics like course creation, prompting, agents, ChatGPT, and YouTube strategy.

💡 In Plain English: Imagine being able to search “what did I say about AI agents?” and instantly getting a structured summary linking back to 12 different videos, docs, and notes — without having to open a single folder.

The Core Limitation for Educators

Here’s where it gets important. The wiki works beautifully — for you. On your desktop. Using your Claude desktop app.

Your students can’t use it.

“My clients and my customers can’t. They don’t have access to this. They don’t have the wiki. They can’t use it when they’re on trainingsites.io or in my campus.” — James

The LLM Wiki is a personal productivity tool. It doesn’t live on your website. There’s no way for a student to query it from inside your community or campus. For your own research and prep work, it’s excellent. But it doesn’t replace a customer-facing knowledge system.

The Two-Layer Model for Teachers

James’s solution is what he already runs at trainingsites.io — and it’s the model educators should build toward:

Layer 1 — Personal (Wiki): Obsidian + Claude on your desktop. Use it for your own research, content planning, and knowledge retrieval. Your Claude desktop app can query the wiki to ground its answers in your real content.

Layer 2 — Customer-Facing (RAG + Published Tutorials): The same content — transcripts, notes, summaries — gets published to your website as tutorials and indexed into a vector database (RAG). James uses Pinecone and Supabase. This powers an AI study bot that lets any student query the full content library by asking a question — not just browsing.

“There’s a place for a wiki, and there’s a place for a RAG, and it really depends on how much content you have, and also who is going to be consuming the content.” — James

At trainingsites.io, this means 360 published tutorials (from 600 videos), each with the video, structured summary, and transcript — all indexed and queryable by students through the AI study bot.

How to Decide Which to Build First

Ask yourself one question:

“Is it your Claude desktop app that’s consuming the content, or is it an end user or a customer that’s going to be consuming the content?” — James

  • If it’s just for you → Start with Obsidian + LLM Wiki. Great for personal knowledge management.
  • If it’s for your students → Publish tutorials + build a RAG. Customers need queryable, web-based content.
  • If it’s both → Build both layers. Wiki for personal use, published tutorials + RAG for your campus.

⚠️ Important: Don’t skip the customer-facing layer because the wiki feels easier to set up. A beautiful personal wiki that your students can’t access doesn’t improve their learning outcomes.

Key Takeaways

  • Karpathy’s LLM Wiki uses Obsidian + Claude to organize your digital exhaust into a visual, queryable local knowledge base — free to set up.
  • The wiki is powerful for personal use (your Claude desktop app) but students can’t access it.
  • The educator-ready solution is a two-layer model: personal wiki for you, published tutorials + RAG for your campus.
  • The deciding question is simple: who’s consuming the content — you, or your students?

Next Steps

James will cover how to install and configure the Obsidian + LLM Wiki setup in a follow-up video. In the meantime, the highest-leverage move is to start collecting your digital exhaust — all those video folders, PDFs, transcripts, and notes — so they’re ready to ingest when you set it up.

If you’re ready to build the customer-facing layer now, start with the Tutorial Library on trainingsites.io to see how the published tutorial + AI study bot model works in practice.

Livestream Details

Tutorial Series

Share This Video

Facebook
Reddit
Twitter
LinkedIn

Creator

Picture of James Maduk

James Maduk

I Build Training & Membership Sites For Your Courses, Coaching & Community. It's a done for you service when you're pressed for time, hate technology, and have no idea how to get started!